Skip to content

james-transient/transient

Repository files navigation

Transient

The trust infrastructure for autonomous agents.


pipx install transient-trace

Autonomous agents can now act without a human in the loop: write code, push to production, call APIs, make purchases, move money. The missing piece is not capability. It is accountability.

Without a governance layer, there is no audit trail, no interception, no recourse. You cannot deploy agents at scale without knowing what they did, what they were authorised to do, and why. Guardrails inside the agent are not governance. They are instructions and instructions live inside the thing you are trying to govern.

Transient is a governance layer that operates outside the agent process. No sandbox. No framework coupling. It intercepts at the process boundary, outside the agent and below the framework. Every governed action produces a tamper-evident receipt before it reaches the system. Receipts are signed with Ed25519 and independently verifiable by any party with no dependency on the issuing system. The agent cannot disable it, route around it, or see past it.

Capability What it provides
Governance Every subprocess call intercepted before execution. Policy evaluated. Action allowed or denied.
Receipts Every decision recorded as an immutable, Ed25519-signed receipt. Independently verifiable.
Memory Blocked actions, patterns, and session context indexed from the governance layer not from agent logs.
Verification Content-producing actions verified against declared intent. Asynchronous. Non-blocking.

None of this requires changes to agent code. None of it can be bypassed by the agent.

Recent updates

  • New compliance attestation command: transient-trace verify now generates a signed OWASP Agentic + EU AI Act attestation from your receipt chain.
  • New TUI dashboard: You can watch your agents with live receipts, alter global configs, and change/add permissions.

Getting started

Transient is the product. transient-trace is the CLI that powers the governance layer. Install it with:

pipx install transient-trace

Install governance shims:

transient-trace wrap install git curl npm pip3 uv sudo --auto-rc
source ~/.zshrc

Boot your agent through Transient. Every session must be launched via transient-trace run. This is how the governance layer wraps the process:

transient-trace run claude
transient-trace --mode strict --packages filesystem,code,privilege,shell run python agent.py

If no mode is specified, Transient runs in audit mode by default. It records everything but never blocks. Transient runs in three modes:

Mode Behaviour
audit Records everything. Never blocks. Default. Use this to observe what your agent does before enforcing rules.
strict Blocks on policy violations before the action executes. Use this for production governance.
permissive Logs violations but never blocks. Use this for testing policy rules.

Set a mode permanently:

transient-trace config set mode strict

View the audit trail:

transient-trace receipts summary
transient-trace receipts list --outcome deny

Governance packages

Package Blocks
filesystem Bulk delete, sensitive paths (~/.ssh, /etc)
code git push to remote, unverified package installs
privilege sudo, su, chmod escalation, user management
shell curl | bash, eval, inline code execution
web SSRF, mutation requests to internal hosts
messaging External broadcast, unknown recipients

Protocol foundation

Transient implements ATP 1.0 the open protocol specification for autonomous agent action governance. Every governed action produces three canonical objects:

Object What it represents
Intent The declared action the agent wants to perform
Decision The governance outcome: allow or deny
Receipt The immutable, cryptographically signed record

Receipts are signed with Ed25519 and independently verifiable by any party with no dependency on the issuing system. ATP is an open protocol. Transient is the reference implementation.

Intended use

Transient is designed for teams building or deploying autonomous agents who need:

  • Infrastructure-level governance that cannot be bypassed by the agent
  • A tamper-evident audit trail for compliance and accountability
  • Memory and verification that operate from the governance layer, not the application layer

How Transient runs

Transient wraps your agent process. You launch your agent through transient-trace run. That is how the governed session is created, the receipt chain is started, and subprocess interception is activated.

transient-trace run python agent.py
transient-trace run claude
transient-trace --mode strict --packages filesystem,code,privilege,shell run python agent.py

Launching an agent directly without transient-trace run means there is no session context and no linked receipt chain. Permanent shims will still intercept specific binaries, but the receipts are unconnected. For full governance, always start through transient-trace run.

Platform support

Platform Status
macOS (Apple Silicon, Intel) Supported
Ubuntu 22.04 LTS Supported
Ubuntu 24.04 LTS Supported
Debian 12+ Supported
RHEL 9 / Rocky 9 / Amazon Linux 2023 Supported
Windows Not supported
Alpine Linux (musl) Not supported

See Linux setup guide for installation instructions on Linux.

Current limitations

Coverage is strongest for terminal-based agents. Agents running through IDE extensions, browser interfaces, or GUI tools may not have full interception. Native Python network calls are not covered at the socket level.

For complete coverage in production environments, pair Transient with a network proxy or OS-level egress filter.

Documentation

Repository structure

transient-trace     → install via pipx (the governance CLI)
src/                → receipt bus — connects Trace receipts to Recall and Intelligence
docs/               → full documentation

transient-trace is the entry point. The receipt bus runs alongside it and feeds governed actions into Recall (memory) and Intelligence (verification) in real time.

Licence

Transient Software License Agreement. © 2026 Transient Intelligence Ltd.

Free during alpha for evaluation and feedback. Free alpha access does not constitute a commitment to provide the software free of charge after the alpha phase ends. Commercial licensing will be introduced at general availability.